MAC2026-track1-salaheiyo1

MAC 2026 — Track 1 — Trained checkpoints & cached scores (test F1_mean = 0.72581)

Companion artifacts for the code package, released as submission_0.72581_checkpoints.zip at https://huggingface.co/salaheiyo/MAC2026-track1-salaheiyo1.

git lfs install
git clone https://huggingface.co/salaheiyo/MAC2026-track1-salaheiyo1
unzip MAC2026-track1-salaheiyo1/submission_0.72581_checkpoints.zip

Two reproduction paths:

Contents

reproduce_from_scores.py    # Path A: scores/ -> prediction.csv, self-checks md5 vs reference
test_order.txt              # the 1138 test vids, in the row order of the scores/pickles
prediction_reference.csv    # the actually-submitted file (target of the md5 check)
scores/                     # per-model TEST softmax scores (the ensemble inputs)
│   test_base_tv.pkl  test_vL_tv.pkl  test_crop_tv.pkl  test_tsf.pkl  test_v1.pkl
checkpoints/                # the 5 trained model weights (weights-only state_dicts)
│   vmae_base_trainval.pth          (VideoMAEv2-Base,  train+val)
│   vmae_large_trainval.pth         (VideoMAEv2-Large, train+val)
│   vmae_base_crop_trainval.pth     (VideoMAEv2-Base + person-crop, train+val)
│   timesformer_ssv2.pt             (TimeSformer-SSv2, HF state_dict)
│   videoswin_small.pth             (VideoSwin-small)
SHA256SUMS                  # integrity

Path A — instant, exact (recommended for verification)

python reproduce_from_scores.py
# -> prediction.csv ; prints "MATCHES SUBMITTED FILE: True"

The 5 weighted (1/1/1/0.7/0.5) softmax tensors are averaged; top-5 = action labels, body = fine->coarse bucketing. This regenerates the submitted file exactly.

Path B — from model weights

Use these checkpoints with the code package to re-dump the score pickles, then run Path A. The checkpoints are weights-only {"state_dict": ...} (mmaction loads them via --checkpoint; no config/optimizer state is needed for inference).

# (in the code package, after set_root.sh + data prep) e.g. for the base model:
sbatch scripts/dump_scores.sbatch  scripts/cfg_vmae_base_ma52_trainval.py \
       /path/to/checkpoints/vmae_base_trainval.pth  /path/to/scores/test_base_tv.pkl
# ... likewise for large / crop / videoswin. TimeSformer:
python scripts/hf_video_finetune.py --model facebook/timesformer-base-finetuned-ssv2 \
       --tag tsf --infer --ckpt /path/to/checkpoints/timesformer_ssv2.pt \
       --infer_split test --out /path/to/scores/test_tsf.pkl --person_crop --clips 2

The provided checkpoints' weights are byte-identical to those that produced scores/, so Path B reproduces the same scores (hence the same prediction.csv).

Notes

• Checkpoints are stripped to weights only (no embedded paths / optimizer / logs).
• scores/ pickles contain only pred_score (+ a basename frame_dir for TimeSformer); no absolute paths or personal data.